Understanding-in-Generation: Reinforcing Generative Capability of Unified Model via Infusing Understanding into Generation

November 11, 2025 · View on GitHub

Yuanhuiyi Lyu¹, Chi-Kit Wong¹, Chenfei Liao¹, Lutao Jiang¹, Xu Zheng¹, Zexin Lu⁴, Linfeng Zhang², Xuming Hu⁴,

¹The Hong Kong University of Science and Technology (Guangzhou)
²Shanghai Jiao Tong University
³The Hong Kong University of Science and Technology
⁴Huawei Hong Kong Research Center

Requirements

Clone the repository:

git clone https://github.com/qc-ly/UiG

cd UiG

Create an environment:

conda create -n UiG python==3.10 -y

conda activate UiG

Install the required packages:

pip install -r requirements.txt

pip install flash_attn==2.5.8 --no-build-isolation

Inference

Please follow official instruction to download the BAGEL-7B-MoT checkpoint and save the checkpoint to ./ckpts.
Generate images from the prompts in ./prompts/test_prompt.txt:
```
bash scripts/infer.sh
```
for slurm:
```
bash scripts/infer_slurm.sh
```

Generate images from input prompts:

python infer.py --prompt_text "A larger person in yellow clothing is partially hidden by a smaller person in a different color."

Evaluation

We follow the official settings of TIIF-Bench and WISE-Bench to evaluate UiG.

The evaluation results are provided in Google Drive

Acknowledgement

Our codes are built on open-source codes, thanks to the following projects:

Thanks for their outstanding works and open-source!

Citation

If you find this repository useful, please consider giving stars ⭐ and citations

@article{lyu2025understanding,
  title={Understanding-in-Generation: Reinforcing Generative Capability of Unified Model via Infusing Understanding into Generation},
  author={Lyu, Yuanhuiyi and Wong, Chi Kit and Liao, Chenfei and Jiang, Lutao and Zheng, Xu and Lu, Zexin and Zhang, Linfeng and Hu, Xuming},
  journal={arXiv preprint arXiv:2509.18639},
  year={2025}
}

Contact

If you have questions, suggestions, and bug reports, please email:

ryan.lyu.mail@gmail.com